EP1836657A2 - Procedes, architecture et dispositif permettant de mettre en oeuvre des systemes a memoire hierarchique et a intelligence machine - Google Patents

Procedes, architecture et dispositif permettant de mettre en oeuvre des systemes a memoire hierarchique et a intelligence machine

Info

Publication number
EP1836657A2
EP1836657A2 EP05853611A EP05853611A EP1836657A2 EP 1836657 A2 EP1836657 A2 EP 1836657A2 EP 05853611 A EP05853611 A EP 05853611A EP 05853611 A EP05853611 A EP 05853611A EP 1836657 A2 EP1836657 A2 EP 1836657A2
Authority
EP
European Patent Office
Prior art keywords
patterns
sequence
module
cortical
processing unit
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP05853611A
Other languages
German (de)
English (en)
Inventor
Jeffrey Hawkins
Dileep George
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Numenta Inc
Original Assignee
Numenta Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Numenta Inc filed Critical Numenta Inc
Publication of EP1836657A2 publication Critical patent/EP1836657A2/fr
Withdrawn legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks

Definitions

  • the present invention relates to the field of memory systems and machine intelligence.
  • the present invention discloses methods, apparatuses, and architecture for implementing memory systems that exhibit machine intelligence.
  • AI Artificial Intelligence
  • Classic artificial intelligence proponents make no attempt to examine or replicate the manner in which the human brain operates.
  • Proponents of classic artificial intelligence attempt to create programs that mimic basic human behaviors or problem solving in a manner that does not incorporate any fundamental understanding of how human brains actually work.
  • People that followed the classic artificial intelligence research school of thought felt that they should not be limited by the particular solution discovered by nature. This school of thought has some resonance since we created flying machines that do not operate the way birds fly and we created fast land travel machines that do not operate the way a cheetah runs.
  • a programmer examines the problem to be solved or the human behavior to be mimicked and then determines an algorithmic solution to the problem. The programmer then codes the algorithmic solution in computer software.
  • Examples of classic artificial intelligence programs include chess playing programs and expert system programs. These programs use an algorithm solution and a set of rules created by a human expert in order to solve complex problems, respectively. However, these programs generally have no ability to learn. These programs can only handle the single problem that was addressed. Nor can such artificial intelligence programs generalize upon the knowledge incorporated into such programs in order to address completely novel input data.
  • Neural network proponents have attempted to create limited intelligent systems by replicating the operation of interconnected neural cells. There is a large body of knowledge that describes how individual neural cells (neurons) operate and how connected neurons interact. Neural network proponents have built systems, known as "neural networks", based upon this knowledge about neuron operation. Thus, neural network systems operate in a manner similar to a set of interconnected neurons. Neural network researchers are therefore often referred to as 'connectionists.' Interneuron connection strengths are known as synaptic weights and are used to store the learned knowledge.
  • a neural network Before being used, a neural network must first be trained with a set of training information.
  • the training information consists of input vectors with associated output vectors that are deemed to be the correct output for the associated input vector.
  • the connections between the various simulated neurons in the neural network are adjusted such that the input vectors generate the associated output vectors (or a close approximation).
  • a neural network is used by presenting a novel input vector to the neural network such that an output vector is generated. With a proper neural network design and adequate training data, the neural network should generate the appropriate output vector for the given input vector. Neural networks have been proven to be useful in some limited applications.
  • neural network systems are relatively primitive. Most neural network systems are simply a three layer structure with a set of input nodes, a set of middle nodes (also known as the 'hidden nodes'), and a set of output nodes. Although neural network systems are able to iearn' in a very simple sense and exhibit a limited ability to generalize, there is clearly no real understanding of the world. Neural network systems merely create an internal function that best maps the training input vectors to the associated training output vectors. Thus, a neural network is only able to generalize in a limited sense by applying the internal function to the novel input vectors.
  • a system may comprise a plurality of individual cortical processing units arranged into a hierarchical structure. Each individual cortical processing unit receives a sequence of patterns as input. Each cortical processing unit processes the received input sequence of patterns and outputs another pattern. As several input sequences are processed by a cortical processing unit, it will therefore generate a sequence of patterns on its output. The sequence of patterns on its output may be passed as an input to one or more cortical processing units in next higher layer of the hierarchy. A lowest layer of cortical processing units may receive sensory input from the outside world. The sensory input also comprises a sequence of patterns.
  • Each individual cortical processing unit includes a memory that is used to help process the sequence of patterns received on the input.
  • the memory may comprise previously encountered sequences of patterns that contain structure.
  • An input sequence of patterns may be deemed to contain structure if that sequence of patterns has been encountered more than once.
  • An individual cortical processing unit can use the stored sequences of patterns containing structure in order to later 'recognize' such structured sequences in the input stream.
  • the cortical processing unit can report such recognition to higher layer cortical processing units.
  • the output to the higher layer cortical processing units may comprise an identifier of the recognized pattern sequence.
  • the output stream containing identifiers of recognized sequences comprises a sequence of patterns itself. That output stream of patterns may comprise an input that will be processed in a similar manner by the cortical processing unit in the next higher cortical processing unit layer.
  • the outputs of more than one lower layer cortical processing unit may be joined together such that the joined outputs form a sequence of patterns that may be processed.
  • a cortical processing unit may also use the memorized sequence of patterns in order to make predictions about the incoming sequences in order to resolve ambiguous information. Specifically, when a cortical processing unit determines that it is currently receiving a known sequence, the cortical processing unit may predict the next pattern to be received in the sequence. The processing unit may use the prediction to interpret the next pattern received. For example, if the input stream into a cortical processing unit is ambiguous, then that cortical processing unit may use the predicted pattern to help identify the next input pattern.
  • the cortical processing unit may further provide the predicted next pattern down to the next lower cortical processing units.
  • the lower cortical processing units may interpret this as a prediction of the next sequence that the lower unit will encounter (and hence output to the higher layer).
  • Such prediction information may comprise information that helps the lower cortical processing unit layers correctly identify the next sequence.
  • Figure 1 illustrates a conceptual diagram of the visual processing portion of the brain concerned with recognizing forms and identifying objects.
  • Figure 2A illustrates a revised hierarchical diagram for the human visual processing system of Figure 1.
  • Figure 2B illustrates an alternate hierarchical diagram for the human visual processing system of Figure 1.
  • Figure 2C illustrates the alternate hierarchical diagram of Figure 2A with a local prediction feedback.
  • Figure 2D illustrates the alternate hierarchical diagram of Figure 2A with a prediction passed from a higher layer to a lower layer.
  • Figure 3A illustrates a grid of pixels in the limited visual world considered by a prototype program machine intelligence program.
  • Figure 3B illustrates the pixel grid of Figure 3A wherein the pixel grid has been divided into a set of sixty-four pixel patches of four by four pixels that will each be examined by an individual processing unit.
  • Figure 3C illustrates the pixel grid of Figure 3B wherein the sixty-four pixel patches have been grouped together in groups of four to form sixteen meta-patches that are examined by the sixteen processing units of the second processing unit layer.
  • Figure 4 illustrates one possible example of the hierarchy of the cortical processing units for the thirty-two by thirty-two grid of pixels illustrated in Figures 3A to 3C.
  • Figure 5A illustrates the hierarchy of the cortical processing units to process information from thirty-two by thirty-two grid of pixels illustrated in Figure 3A in a prototype program.
  • Figure 5B illustrates the hierarchy of Figure 5A with feedback information that provides context.
  • Figure 6 illustrates an alternate view of the cortical processing unit hierarchy of Figure 5A wherein the cortical processing units are each labeled with the addressing system set forth with reference to Figures 3B and 3C.
  • Figure 7 illustrates a Bayes belief net for the cortical processing unit hierarchy of Figure 5 A.
  • Figure 8 illustrates ninety training shapes that were used to train the prototype program.
  • Figure 10 illustrates some result data from three different line drawing tests using a prototype program.
  • the human brain consists of billions of interconnected nerve cells (neurons). Like a computer system, the neurons of the human brain are connected to both input units and output units.
  • the input units consist of the senses connected to our brain such as hearing, sight, smell, touch, and taste.
  • the output units consist of controlled muscles that allow us to perform complex actions with our bodies and allow us to communicate by speaking. But beyond that one similarity, computers and human brains are extremely different.
  • a computer has a central processor that sequentially executes a set of instructions (with associated parameter information) retrieved from a flatly addressed memory system. In a human brain, there is no central processor. Instead, there are billions of neurons interconnected in some parallel manner wherein all the neurons operate concurrently.
  • the human brain has been studied by scientists for over a hundred years.
  • the teachings of the present invention combine a number of the discoveries about the human brain in order to create an overall theory on how the human brain operates.
  • the operational theories mainly focus on the brain's neocortex since only mammals, the class of animals generally regarded to be the most intelligent, have a neocortex.
  • a full description of the overall theory of operation of the brain's neocortex can be found in the book "On Intelligence” by Jeff Hawkins, Times Books, which is hereby incorporated by reference in its entirety.
  • these teachings can then be implemented with a machine in order to create artificial intelligence.
  • the methods of the present invention may be implemented with software on a computer system.
  • the teachings can be implemented in many other formats such as directly in an integrated circuit.
  • the human brain is protected within the confines of the strong bone that consists of the human skull. There is no direct contact between the human brain and the outside world. To learn about the outside world, the human brain has millions of connections to sensory nerve cells located all over the human body. Millions of nerve cells on the surface of the human body provide the input of touch. Millions of rods and cones on the retina of the human eye detect light and transmit information about the detected light along the optic nerve to the brain. Thousands of sensor cells in the human ear detect different sound frequencies and transmit the detected sound information along the auditory nerve to the brain. All of this sensory information is received along different nerve pathways to the brain. The brain must then interpret all of this information to form an understanding of the outside world.
  • All of the different sensory input signals that are transmitted to the brain have at least two dimensions: a spatial dimension and a time dimension.
  • the spatial dimension constitutes a collection of related sensory signals that arrive at the brain at the same time.
  • the collective activities on related nerve fibers that are received by the brain at the same time are deemed to be a "pattern" of information in the spatial dimension.
  • different patterns are received by the brain from each collection of related nerve fibers.
  • the consecutive patterns received by the brain along related nerve fibers constitute a 'sequence of patterns' wherein consecutive patterns are closely related in time. An example can easily be provided with reference to the human sense of vision.
  • All of the input signals received by the brain from the optic nerve at a particular instant in time represent a pattern of visual information in the spatial dimension.
  • the pattern in a spatial dimension and a sequence of patterns in the time dimension model applies to all of our different senses.
  • different hairs in our ears are activated by different frequencies of sound.
  • the collection of all of the different frequency signals transmitted from the ear to the brain along the auditory nerve represents a pattern of auditory information.
  • We hear different sounds such that sequence of auditory patterns is transmitted along the auditory nerve to the brain.
  • Touch information has a wide variety of spatial dimension aspects including pressure, temperature, pain, and location. All of this touch information that the brain receives concurrently is a pattern of touch information in the spatial dimension. Over time, the body receives a sequence of patterns of touch information.
  • Much of the artificial intelligence research using neural networks concerns the use of neural networks attempting to learn pattern information such that the neural network can later recognize similar patterns within novel input data.
  • most of the neural network based pattern recognition research has been performed using spatial patterns with no consideration of time. For example, during a learning phase, a neural network may be presented with a collection of static images, one at a time, such as an apple, a boat, a tree, a banana, etc. The neural network is 'taught' what each presented pattern represents in order to train the neural network on different object patterns for object recognition. Later, novel image patterns are presented to the neural network such that the neural network attempts to identify an object in the novel image. There has been some success with such pattern recognition neural networks, but the success has been limited.
  • Human brains do not receive information in such a static manner wherein a set of completely different images are presented with no respect to time. Instead, human brains receive information as a continuous sequence of patterns from the different sensory organs. For example, our eyes receive an ever-changing sequence of images that are transmitted to the brain as a sequence of patterns on the optic nerve. Our brains learn to group different sequential images of the same object together by observing continuous motions in this world.
  • the sensory information in the sequence of patterns is generally highly correlated over time. For example, when a person looks out a window, much of the visual information (such as buildings, tree trunks, the ground, etc.) remains static. Some of the visual information (such as leaves blowing in the wind, a bird flying in the sky, a squirrel running across the grass, etc.) changes over time. But even the changing information is highly correlated in time since the leaves blowing in the wind do not move significantly, the bird travels in a straight line, and the squirrel moves around only on the ground in the same area.
  • the visual information such as buildings, tree trunks, the ground, etc.
  • the system of the present invention proposes that the consideration of time such that a sequence of patterns is examined to be an important aspect of in the ability to learn.
  • the system of the present invention uses sequences of patterns generated over time as the stream of source information to learn from.
  • the Neocortex in the Brain Uses a Single 'Cortical Algorithm'
  • the human brain is comprised of many different parts including the cerebellum, the basal ganglia, the hypothalamus, the brain stem and the neocortex.
  • the main area of the human brain responsible for abstract thought and substantial learning capability is the neocortex. Only the animals that exhibit the best capacity for learning (mammals) have a neocortex structure in their brains.
  • neocortex Different regions of the neocortex are generally used to address different problems presented to the neocortex.
  • One region of the neocortex is used to process grammar; other regions are used for processing sounds; multiple regions are dedicated to processing visual information; and so on.
  • scientists have attempted to map out the different functions performed by the different regions of the neocortex.
  • the current brain function mappings of neocortex functions are relatively complex. And it has been found that the functional mappings of the neocortex only tend to vary by limited amounts from person to person.
  • neocortex The relatively consistent mapping of neocortex functions has lead many brain researchers to assume that each area of the brain processes the information that it receives and solves problems in different ways.
  • a neuroscientist named Vernon Mountcastle made a novel assertion. He noted that the entire neocortex had a remarkably uniform appearance and structure. The regions of neocortex that handle auditory input look the same as the regions that handle touch, which look like the regions that control muscles, which look like the regions that handle vision, which look like the regions that generate language, and so on. Mountcastle asserted that since all these different regions of the neocortex appeared uniform, then perhaps all of these different regions of the neocortex performed the same basic operation.
  • the same plasticity has been shown to exist in human brains. For example, if the region of the neocortex that is usually dedicated to processing grammar (known as Broca's area) is damaged in the brain of a young child that has not yet learned to speak, then the child's damaged brain will adjust itself accordingly in order to adapt. Specifically, the child's brain will tend to use a different region of the brain for processing grammar (such as the same area on the other side) such that the child with a damaged brain will easily learn to speak. Many other examples of such human brain plasticity can be cited. Adults that are born deaf will use areas that are normally used to process sounds for processing visual information. Blind people will often use the region of the rear region of the brain normally used to process vision in order to learn to read Braille.
  • Figure 1 illustrates a simplified conceptual diagram that is commonly used by brain researchers to describe the visual processing portion of the human cortex concerned with recognizing forms and identifying objects in the visual information stream.
  • raw visual information enters the Vl region 110.
  • the Vl region 110 performs the initial processing of visual pattern information received from the optic nerve.
  • the cells in the Vl region 110 are divided into cell groups wherein each cell group analyzes the information from a small area of the retina.
  • all the neural cells in a cell group that process the same sequence of input patterns or an artificial device that emulates a cell group that processing the same sequence of patterns will be referred to as a "cortical processing unit” or a "processing unit.”
  • a cortical processing unit a cell group that processes the information contained in sequence of input patterns directly from a small area of the retina may be referred to as a cortical processing unit.
  • the information generated by the processing units in the Vl region 110 is then passed to the V2 region 120.
  • the V2 region 120 performs processing on the information received from the Vl region 110 and then generates processed output information that is passed to the V4 region 130.
  • the V4 region 130 processes the information received from the V2 region 120 and passes processed information to IT region 140.
  • the IT region 140 receives information generated by the V4 region 130.
  • the brain is known have been able to discern objects in the person's visual field. For example, certain cells will be activated if a particular face is visible anywhere in the person's visual field.
  • raw visual sensory information is received at the bottom and then progressively processed by a series of visual processing regions.
  • the visual processing system is known to recognize and identify objects.
  • the visual processing portion of brain concerned with recognizing forms illustrated in Figure 1 has a vertical hierarchy of successive processing until a complex result (an identified object) is achieved at the top.
  • Vl region 110 the cells in the lower layers of the hierarchy (such as the Vl region 110) are switching (activating and deactivating) very rapidly. This is understandable since our eyes are constantly moving such that different visual information is being presented to the Vl region 110 of the cortex all the time.
  • the Vl region 110 is also the largest cortex region of the four cortex regions conceptual illustrated in Figure 1. Furthermore, experiments have shown that each cell group in the Vl region 110 is only receptive to a very small area of the retina. Thus, in the Vl region 110 there are many individual small cell groups responsible for processing information from each area of the retina.
  • the activity of the cells at each higher hierarchical layer become more stable than the lower hierarchical layers.
  • cells respond more and more to objects in the world as opposed to simple visual constructs.
  • the receptive visual fields of the cell groups in each layer become larger as one ascends the hierarchy. That is, cell groups in the low Vl region 110 only process information from a small area of the retina whereas the IT region 140 processes information from the entire visual field.
  • each cell group in a layer above the first layer processes information from more than one cell group from the layer below. This can be seen since a face can appear in any region of a person's visual field and the person still recognizes it as a face such that the IT region 140 must process the entire visual field.
  • the cell groups in each ascending layer process a larger and larger visual field such that each cell group in regions above the first layer (Vl region 110) processes the output of many cell groups from regions in the lower layers.
  • the IT region 140 processes the entire visual field since objects can be recognized in any part of the visual field.
  • each cortical processing unit receives sequences of patterns that must be processed. Combining these teachings, one may deduce that each layer of the hierarchy must comprise cortical processing units that receive input in the form of a sequence of patterns but output a sequence of patterns since those output patterns will become the inputs to the cortical processing units in the next higher layer.
  • Figure 2A illustrates a block diagram one embodiment of a revised hierarchical diagram for the human visual processing system wherein each rectangular block represents a cortical processing unit.
  • the hierarchical organization may not always consist of progressively fewer cortical processing units on each ascending layer.
  • Figure 2B illustrates an embodiment of a cortical processing unit hierarchy wherein the third layer from the bottom has more cortical processing units than the second layer from the bottom.
  • FIG. 2A illustrates a very simplified version of the proposed hierarchical architecture of the visual processing portion of the human neocortex concerned with recognizing forms and identifying objects.
  • each layer of the hierarchy comprises one or more cortical processing units wherein each cortical processing unit is represented by a rectangular box.
  • Figure 2A is just provided for conceptual explanation purposes. In reality, there are certainly many more layers and there are certainly many more processing units on each layer. With the proper determination of the single cortical algorithm that drives each processing unit, an intelligent machine may be constructed using the hierarchical architecture illustrated in Figure 2A in order to perform visual object recognition.
  • the Cortical Algorithm Stores Structure in Memory for Later Recognition
  • the sensory input information received at the bottom of the hierarchy consists of sequences of patterns from the sensory organs.
  • each cortical algorithm processing unit would likely also output patterns (that, over time, are sequences of patterns) such that those patterns may be used as the inputs for the next higher layer of cortical processing units in the hierarchy.
  • the challenge is to determine what algorithm is used internally in each cortical processing unit to process the incoming sequence of patterns and generate an outgoing sequence of patterns.
  • each cortical processing must include memory for storing information. Since each cortical processing unit receives an input sequence of patterns, has a memory, and outputs information; then that memory must be used by the cortical processing unit to help process the input sequence of patterns in order to create the output sequence of patterns.
  • An invariant representation is an internal representation of an object (a physical object, a word, a song, etc.) in the brain that allows the brain to later recognize the same object in just about any different external condition.
  • a brain's invariant representation of a particular person allows the brain to later identify that person in any lighting condition, from any viewing angle, and from any distance provided that there is enough information in the sensory input stream to identify that person.
  • the system of the present invention proposes that an invariant representation of an object can be constructed by using a hierarchy of cortical processing units.
  • the invariant representation is formed by identifying invariant structure within the constantly changing sequence of patterns and storing that invariant structure so that it may later be recognized. Every cortical processing unit performs this process of identifying, storing, and later recognizing invariant structure at every level in the hierarchy.
  • the overall invariant representation of an object comprises all of the invariant structures associated with that object that have been stored in a set of hierarchically arranged cortical processing units.
  • an invariant representation includes invariant structures at many different levels. All of these different levels invariant structures (including the entire invariant representation structure) can be used within the invariant representations of other objects.
  • the neocortex must create these invariant representations of objects using the neocortex's only source of outside information, the constantly changing sequence of patterns from the various sensory organs.
  • the neocortex must create an invariant representation of an object by identifying invariant structures within the constantly changing sequence of patterns.
  • the memory in each cortical processing unit must be used to store invariant structure that has been located in the incoming sequences of patterns. In this manner, the stored invariant structures may later be recognized in the incoming sequence of patterns in order to recognize objects that include that invariant structure.
  • a non- structured sequence of patterns is generally represented as a completely random stream of patterns. So a structured stream of patterns is not a random stream of patterns. But how can a cortical processing unit that has not had any learning identify a sequence of patterns that is not random?
  • One possible method is to have the cortical processing unit look for a repeated patterns or repeated sequences of patterns. Repeated patterns and repeated sequences of patterns are unlikely to be random noise. Furthermore, the more times that a pattern or sequence of patterns repeats, the less likely that the repeated pattern or the repeated sequence of patterns is random. Thus, repeated patterns and repeated sequences of patterns are likely to be part of an invariant structure.
  • a number of other methods may also be used to identify invariant structures with the stream of patterns. For example, ordered progressive pattern changes such as 1, 2, 3, 4, and so on are also mostly likely to be invariant structure.
  • each cortical processing unit may search for repeated patterns or repeated sequences of patterns in the input stream in order to identify invariant structures contained within the input stream.
  • the cortical processing unit may then store those repeated patterns and repeated sequences of patterns in memory for later recognition.
  • a cortical processing unit may only memorize the most frequently repeated patterns or sequences of patterns. This limited number of repeated patterns or sequences of patterns stored in memory may be considered as a set of 'categories' that are used to classify incoming patterns or sequences of patterns. Later, when a previously seen and memorized pattern or sequences of pattern (a learned invariant structure) is encountered again, the cortical processing unit can inform the next higher cortical layer about the identified invariant structure found in the input stream.
  • a cortical processing unit may pass up a sequence identifier (or name) for the recognized sequence of patterns to the next higher layer in the hierarchy. As long as the incoming sequence of patterns continues to match the memorized sequence of patterns, the cortical processing unit will continue to output the same sequence identifier for that recognized sequence of patterns. Since the same sequence identifier is being output during the entire pattern sequence, one can see that higher layers will have processing units that switch less frequently than lower processing units.
  • the cortical processing unit may output an identifier to the cortical processing unit in the next higher layer that indicates that nothing was recognized in the current input stream.
  • the cortical processing unit may output an identifier that indicates that there was an error between a predicted pattern and the actual pattern received.
  • the sequence identifier itself is a pattern on the cortical processing unit's output. And the output pattern (the sequence identifier) becomes the input to the next higher cortical processing unit. Over time, a cortical processing unit will output different sequence identifiers as different sequences are recognized such that a sequence of patterns (a sequence of sequence identifiers) is output from the cortical processing unit.
  • the cortical processing unit in that next higher layer can then apply the same cortical algorithm to its input stream(s) from one or more processing units in lower layers of the hierarchy. Specifically, the cortical processing unit in that next higher layer can examine the input stream for repeated patterns or sequences of patterns and store those repeated patterns and/or repeated sequences of patterns. When the cortical processing unit recognizes a stored sequence of patterns in the input stream, that cortical processing unit can inform a cortical processing unit in the next higher layer with a sequence identifier for the recognized sequence. In this manner, the hierarchy of cortical processing units (as illustrated in Figure 2A) is able to identify structures within structures.
  • the hierarchical structure promotes efficient usage of memory since simple structures in lower hierarchical layers can be combined in many different manners to create more complex structures in higher layers.
  • low level cortical processing units may recognize simple invariant structures such as lines of different orientations (vertical, horizontal, 45 degree, etc.).
  • higher level cortical processing units may recognize more complex structures by combining together different permutations of the line orientations in the lower-level structures.
  • a higher level cortical processing unit may recognize a rectangle as a first pair of parallel lines that are coupled at each end by a second pair of parallel lines with an orientation that is orthogonal to the first pair of lines.
  • a square could be recognized as first pair of parallel lines that are coupled at each end by a second pair of parallel lines having an orientation that is orthogonal to the first pair of lines and a length equal to the first pair of parallel lines. Note that the same low level structures are used in both higher level object definitions.
  • the rectangle structure itself may be primitive structure used to identify more complex objects such as a house, a billboard, a television, etc.
  • the cortical processing units in each successive layer become more stable since the identified sequence of patterns will change less frequently.
  • the information becomes very stable.
  • a high layer such as IT region in the human neocortex's system for identifying visual objects
  • a consistent visual pattern such as a human face can be consistently recognized even when it moves about the person's field of view thus creating very different visual input information that causes rapid switching in lower regions such as the Vl region.
  • studies have shown that certain cells in the high levels such as the IT region remain activated as long as there is a human face within the field of view even when the face moves about the field of view.
  • the Cortical Algorithm makess Predictions Using Its Stored Memories
  • intelligence is measured by the ability to make useful predictions.
  • Intelligence allows humans (and other mammals) to use accumulated knowledge about the world in order to make predictions about the world. Simply identifying invariant structure in the surrounding world alone is not very useful. However, using that accumulated knowledge about the world in order to make accurate predictions about that world is a very useful adaptation for an animal struggling to survive in the world.
  • an entity can make accurate predictions about an object or situation, then that entity has intelligence such that the entity can be said to 'understand' that object or situation.
  • an animal that is able to identify a previously encountered situation and then make accurate predictions about that situation will be better able to survive and reproduce.
  • a gazelle that previously encountered a lion catching and eating another gazelle may learn from that encounter such that it can predict that hanging around lions may get oneself killed. Based upon that prediction, the gazelle may opt for behaving in a manner that prevents the undesirable prediction from coming true . . . Run!
  • an animal with the ability to predict may adapt its behavior to cause favorable predictions to come true. If an animal previously encountered a watering hole by the patch of five trees, then the animal may predict that going to those five trees would again provide a watering hole. Therefore, natural selection favored animals with a neocortex well suited for making accurate predictions.
  • the neocortex likely uses a single cortical algorithm in all the cortical processing units arranged in a hierarchy to address many different problems.
  • both high level cortical processing units and low level cortical processing units make predictions.
  • Very high level cortical processing units may make sophisticated decisions such as those presented in the previous paragraph that helped an animal survive.
  • very low level cortical processing units constantly make very simple predictions.
  • a person's neocortex constantly makes many of these low level predictions with out that person being aware of those predictions. Those low level predictions are generally only of interest to the surrounding low level cortical processing units.
  • failed low level prediction that escalates into our consciousness that almost everyone has experienced is that sudden feeling of panic that you feel when you are walking up stairs and your foot passes through the plane where you predicted the next step should be.
  • Your brain notes that reality does match the prediction such that the problem rapidly escalates up the cortical hierarchy. Since no intermediate cortical processing unit is able to resolve this failed prediction, the failed prediction escalates all the way up into your consciousness in order to bring your attention to the problem.
  • Our brain can adjust to mismatches between sensory input and predicted reality in many different ways. If the sensory perception is deemed to be accurate but the prediction is not then the error indicates that the brain's model of the world is not accurate. The brain may adjust its model of reality by learning. Thus, in the preceding example, you may learn that the set of stairs has fewer steps than you previously expected.
  • each cortical processing unit stores sequences of patterns in memory (Ideally, sequences of pattern that represent structure in the world) for later recognition. If the incoming patterns match a particular memorized sequence, then the cortical processing unit 'recognizes' that particular sequence and informs the coupled processing units in higher layers that a known sequence is being observed. However, this sequence recognition can also be used to make a prediction.
  • a cortical processing unit may determine the next pattern that the processing unit expects to receive by identifying the next pattern in the recognized sequence of patterns. This next expected pattern constitutes a prediction by the cortical processing unit. Such predictions occur all the time in our brains at every level of the cortical processing unit hierarchy.
  • the next pattern prediction made by a cortical processing unit may be used as an input to that processing unit to be processed.
  • Figure 2C illustrates the cortical hierarchy of Figure 2A wherein cortical processing unit 231 has made a prediction of the next expected pattern (represented by the dashed line) and that prediction that is passed back into cortical processing unit 231 as input information.
  • the predicted pattern value may be delayed before being introduced into the input stream.
  • On Intelligence it was suggested that the thalamus may be used to perform this function in the human brain.
  • the cortical processing unit may compare its prediction of the next pattern to be received from lower processing units with the actual pattern received. In such a manner, the cortical processing unit may be able to use the pattern prediction to resolve ambiguous information. For example, if the next pattern is very close to the expected pattern, then the cortical processing unit may deem that the next pattern was really meant to be the predicted pattern but it may have been distorted by noise. If the cortical unit presumes that the predicted pattern was really what was meant, then the cortical processing unit will deem that it is still within the recognized sequence. Thus, a pattern sequence that is very similar to a predicted pattern sequence may effectively be changed into that predicted sequence.
  • a cortical processing unit may create a perceived reality by combining an actual real but ambiguous input (a closely matching pattern) with a known model of reality (the expected pattern) to create a perception of reality. If the expected pattern is significantly different than the expected pattern, then the cortical unit may try to recognize a different known sequence of patterns, signal an error to higher units, signal the error to lower units, or act in another manner.
  • An auto-associative memory is a memory that produces a full clean version of known pattern or sequence of patterns when give a noisy or partial version of that known pattern or sequence of patterns.
  • An auto-associative memory a group of similar input patterns or sequences of patterns will all retrieve and output the same output pattern or sequence of patterns. This is a form of generalization that humans can do effortlessly.
  • most auto-associative memory systems will fail if the input pattern is shifted, rotated, scaled, or otherwise distorted in a significant manner. Human memories are often linked together by their own content like auto associative memories.
  • Auto-associative memories can be used to store sequences of patterns. To store a sequence of patterns in an auto-associative memory, one can associate the current input pattern with the previous input pattern, then associate the next input pattern with the current input pattern, and so on. Then, the output of the auto-associative memory is coupled to the input of the auto-associative memory. In this manner, when the auto-associative memory is primed with a spatial pattern that is associated with the first pattern in the sequence, the entire sequence will be recalled.
  • Auto-associative memories may be used to construct primitive versions of the cortical hierarchy of the present invention.
  • auto-associative memory devices may be arranged in a hierarchical arrangement wherein the outputs from low-level auto associative memories are coupled to the inputs of higher auto-associative memories with some additional circuitry.
  • the auto-associative memories store, recognize, and recall the pattern sequences mentioned earlier.
  • the additional circuitry is needed to transform the output sequences into names of output sequences that will be provided to the next higher level of auto-associative memories.
  • next pattern prediction may also be passed between different cortical processing units.
  • Figure 2A illustrates the different hierarchical layers with information flowing both upwards and information flowing downwards.
  • each individual cortical processing unit attempts to identify invariant structure in its input stream.
  • information about that recognition an identifier for the recognized sequence
  • the downward information flow may include next pattern predictions made by higher cortical processing units in the hierarchy to lower higher cortical processing units.
  • FIG. 2D explicitly illustrates a conceptual example of such a downward flow of prediction information.
  • cortical processing unit 231 has made a prediction (represented by a dashed line) that is passed to the two cortical processing units (221 and 222) in a lower cortical layer that are coupled to cortical processing unit 231.
  • Each cortical processing unit memorizes sequences of patterns that the cortical processing unit receives on its input. Those inputs are actually identifiers of recognized sequences from the cortical processing units in the lower layer.
  • the cortical processing unit may predict the next pattern in the sequence it is currently receiving. If that predicted pattern is passed to the cortical processing unit in the hierarchical layer below, that predicted pattern represents the next sequence of patterns that is expected from the lower cortical processing unit.
  • the lower cortical processing unit receives a prediction about the next expected sequence of patterns it will likely encounter.
  • Such a prediction about the next expected sequence may help a cortical processing unit resolve ambiguity on a sequence of patterns level.
  • a cortical processing unit may not be able to match a new incoming sequence of patterns to any of the stored sequences of patterns. Or worse, the incoming sequence of patterns may closely resemble two different stored sequences of patterns.
  • the cortical processing unit may consider the prediction information received from a higher cortical processing unit.
  • a cortical processing unit may use that prediction from the higher cortical processing unit to resolve such ambiguous cases and thus select sequence B.
  • This complex prediction system that creates predictions both within individual cortical processing units and between coupled cortical processing units creates a sophisticated prediction system that operates throughout the entire hierarchy in a coordinated manner. Specifically, every cortical processing unit in the hierarchy may use its own memory to predict the next pattern to be received in the currently recognized sequence of patterns. And every cortical processing unit in the hierarchy may pass that predicted next pattern down to the lower connected cortical processing units. A cortical processing unit may use a prediction received from a higher cortical processing unit to predict the next sequence of patterns that will be received.
  • each pattern or pattern sequence prediction matches the received input, things run smoothly. And predictions may be used to resolve ambiguous patterns or pattern sequences. But when a prediction is significantly wrong, an error signal or the unexpected input is propagated up the cortical hierarchy. If a cortical processing unit higher up in the hierarchy is able to resolve the problem, then correction information will flow from that cortical processing unit back down the hierarchy. If a problem propagates all the way to the 'top' of a hierarchy, then some new learning may need to occur in order to address the novel situation in the future.
  • the collective memory of invariant structures learned by the hierarchy of cortical processing units comprises an invariant representative model of the surrounding world.
  • our neocortex By combining incoming sequences of patterns from sensory organs along with sequence and pattern predictions that are constantly made by all the cortical processing units using the invariant representative model of the surrounding world stored in the memory, our neocortex generates a full perception of reality.
  • our understanding of 'reality' is more than simply what we can detect with our senses at any particular moment. Instead, our perception of reality is created with both sensed information (from our sensory organs) and information recalled from memory (our model of the world), (hi fact, the majority of our perception of reality may come from memory.)
  • the brain combines the sensed visual information with predicted visual information such that the hole from the blind spot is 'filled in' to create the full visual perception. Specifically, the brain fills in the hole with information from the surrounding visual scene and makes intelligent predictions to fill in the hole. If the blind spot hole is within the pattern of a carpet, then that carpet pattern will be used to fill the hole. If a straight line passes through the blind spot, then a straight line segment will be filled into the blind spot in order to complete the straight line. Since our brains are very good at making predictions about our surroundings due to years of learning, the predicted information appears to be just as real as the actual sensed information. Thus, you really cannot always believe what you "see”.
  • Cortical Algorithm Makes Predictions from Invariant Representations and Input Data
  • Feedback or information flowing down the cortical hierarchy serves a number of functions.
  • the feedback information can be used to resolve ambiguous input data. But the feedback data is also used for other very important functions.
  • the feedback information may be required for learning.
  • every cortical processing unit and region of neocortex learns sequences of patterns, that cortical processing unit or region of neocortex learns the sequences in a particular context that is provided by the next higher cortical processing unit or region of neocortex.
  • the context specifies a state of the higher cortical processing unit or region of neocortex.
  • the context may be specified as the current element of a sequence being experienced by that higher cortical processing unit or region of neocortex.
  • This context feedback information provides a means for the cortical processing unit or region of neocortex to properly group its input patterns together.
  • Predictions based on invariant representations occur when a high level pattern moves down the hierarchy towards the input. Such predictions occur in two different manners.
  • a first manner occurs when the system predicts something that has not yet happened. For example, our neocortex allows us to predict (and hear) the next note of a song before it has happened.
  • the other invariant representation type of prediction occurs when a system fills in missing information in something that is happening now. For example, our neocortex fills in missing or occluded parts of a visual image as previously described with reference to the blind spot.
  • the system has a challenge in making the desired specific prediction.
  • the challenge presented is that a pattern high up in the cortical hierarchy is in an invariant form.
  • the invariant form lacks the details to generate a prediction that is precise and relevant to the current situation. But ideally, the system should generate as precise of a prediction as is possible. Thus, the invariant representations must be combined with some specific feed forward information to make specific predictions.
  • invariant representations must be combined with some feed forward information in order to generate specific predictions.
  • the general method to do this is as follows. As feed forward inputs arrive at a cortical processing unit or a region in a memory hierarchy from the units below, those feed forward inputs may be combined to form a spatial pattern. As part of the learning process, this spatial pattern must be stored as one type of example of whatever higher level context information is being fed down to this cortical processing unit or region. As different feed forward inputs (combined into spatial patterns) arrive, the feed forward inputs (spatial patterns) are stored as part of different higher-level contexts that are provided at that time.
  • a cortical unit or neocortex region may experience the following learning: pattern A (feed forward input pattern) is a dog (context), pattern B is a dog, pattern C is a cat and pattern D is a cat.
  • pattern A, B, C, and D patterns are the components of invariant representations that are learned for future reference.
  • the stored patterns are later used during recognition. For example, when feedback information to a cortical unit or neocortex region specifies an expectation of seeing a dog, then the cortical units or neocortex region partially invokes all the stored patterns that were stored in the context of "dog". In the example from the previous paragraph, the cortical units or neocortex region will partially invoke patterns A and B since those patterns were stored in the context of "dog". To complete the prediction process, the partially invoked invariant representation must be combined with current information. Thus, partial feed forward input to the cortical units or neocortex region will preferentially match pattern A or pattern B such that the combination of the feed forward input and its matching pattern results in a precise prediction.
  • the prediction is a specific prediction (pattern A or pattern B) created by combining the feed forward input to the known examples of the "dog" context.
  • the book "On Intelligence” sets forth the details of one mechanism that may be used within the human neocortex. This mechanism uses sparse distributed spatial patterns.
  • layer 2 cells in the six layer columns of neocortex are a superset of all relevant invariant representation patterns. Feedback information from above activates an invariant representation in layer 2 and the activated layer 2 cells signal layer 6 cells in the same column. Feed forward input information from below representing the current situation is provided to layer 6 of the columns relevant to the feed forward information. The intersection of feedback created layer 2 cell activity that signals layer 6 and feed forward activity in layer 6 determines the specific prediction.
  • the system must have multiple different representations of multiple different objects (the different contexts). Each of the multiple different representations must be associated with a specific type of feed forward (or input) information such that the feed forward information will partially activate the representation. Feed back information that specifies a particular context (object) must be able to partially activate the different representations associated with that context (object). Any representation that is partially activated by feedback information and partially activated by feed forward information is considered fully activated.
  • a specific prediction can be made by specifying a particular context with feedback information (to partially activate a relevant set of representations) and specifying current feed forward input information (to partially activate representations related to the feed forward information) such that a fully activated representation is generated as a prediction.
  • the prediction may then flow down the hierarchy.
  • the present invention proposes using aspects of the architecture and the operation theory of the human brain set forth in the previous sections as a blueprint for creating sophisticated memory systems and intelligent machines.
  • the human neocortex By attempting to emulate important aspects deduced from the sophisticated intelligence system honed by evolution (the human neocortex), significant advances in the field of artificial intelligence may be achieved. This approach avoids the pitfalls of the current main two approaches of artificial intelligence.
  • the subsections of the previous section set forth the main principles that have been derived from the mammal neocortex. Specifically:
  • Sensory Input Comprises Sequences of Patterns
  • the Neocortex Uses a Single 'Cortical Algorithm' for many different problems 3.
  • the Neocortex Is Organized in a Hierarchical manner
  • An intelligent system may comprise a plurality of individual cortical processing units arranged into a hierarchical structure. Each individual cortical processing unit receives a sequence of patterns as input. Each cortical processing unit processes the received input sequence of patterns and outputs another pattern. As several input sequences are processed by a cortical processing unit, it will generate a processed sequence of patterns on its output that will be passed to cortical processing units in higher layers.
  • Each individual cortical processing unit includes a memory that is used to help process the input sequence of patterns.
  • the memory may comprise previously encountered sequences of patterns that contain invariant structure.
  • An input sequence of patterns may be deemed to contain structure if that sequence of patterns is repeated.
  • An individual cortical processing unit can use the stored sequences of patterns in order to 'recognize' such sequences in the input stream.
  • the cortical processing unit can report such recognition to higher layer cortical processing units.
  • the output to the higher layer may comprise an identifier of the recognized sequence of patterns in the incoming sequence of patterns.
  • the output stream containing identifiers of recognized sequences comprises a sequence of patterns itself. That output stream of patterns may comprise the input to a cortical processing unit in the next higher cortical processing unit layer.
  • a cortical processing unit may also use the memorized sequence of patterns in order to make predictions about the incoming sequences in order to resolve ambiguous information. Specifically, when a cortical processing unit determines that it is currently receiving a known sequence, the cortical processing unit may predict the next pattern to be received in the sequence. The processing unit may use the prediction to interpret the next pattern received. For example, if the input stream into a cortical processing unit is ambiguous, then that cortical processing unit may use the predicted pattern to help identify the next input pattern.
  • the cortical processing unit may further provide the predicted next pattern down to the next lower cortical processing units.
  • the lower cortical processing units may interpret this as a prediction of the next sequence that it will output.
  • Such prediction information may comprise information that helps the lower cortical processing unit layers correctly identify the next sequence. .
  • the prediction information may also be considered as useful predictions of what will occur in the outside world. And if an incoming pattern significantly differs from a predicted pattern or an incoming sequence of patterns significantly differs from the predicted sequence of patterns, then there is a prediction failure.
  • the occurrence of a prediction failure may be transmitted to a cortical processing unit in a higher layer.
  • a cortical processing unit in a higher layer may be able to resolve the error.
  • a detected error may be resolved in a number of different ways. One way to resolve the error is to adjust the cortical hierarchy's model of the world by storing a new sequence of patterns.
  • the previous sections have described a system that has a hierarchy that stores sequences and makes predictions of future events. However, like the human visual system, the system can be used to solve spatial recognition problems.
  • This section presents one example embodiment of a prototype program that uses the teachings of the previous sections to recognize visual object patterns by first learning about the object patterns.
  • the principles of operation and the architecture are set forth with sufficient detail such that many different systems that incorporate the teachings of this document may be created.
  • the prototype pattern recognition program is constructed of a hierarchy of cortical processing units that analyzes input training information consisting of line drawings in order to identify structures within that training information.
  • the prototype program is also informed about the names of objects in some of the training information such that the prototype program can associate particular learned structures with particular object names. After the prototype program completes its learning phase, the prototype program can then accurately recognize objects in novel input data using the learned information.
  • the hierarchical processing units of the prototype program receive pattern information that is presented as a continuous stream in the form of animated movies.
  • sequences of patterns are presented to and learned by the prototype program.
  • the prototype program uses the learned information to create probabilistic models about the outside world.
  • the probabilistic models are then used to recognize objects in novel pattern data using Bayesian belief propagation as is well-known in the art. Explicit details on Bayesian belief propagation can be found in the book "Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference" by Judea Pearl” from Morgan Kaufman Publishers, Inc.
  • other implementations that incorporate the teachings of the present invention may process the learned information in different manners and may use other methods of inferring a final result.
  • the Visual Pattern Recognition Problem Addressed The Visual Pattern Recognition Problem Addressed
  • the visual pattern recognition problem addressed by the prototype program implements the hierarchical architecture of the present invention with computer software. To keep the size of the software, the amount of memory required, and processing requirements of the prototype program under control and to quickly create a functioning prototype, some simplifications were made from an ideal implementation of the present invention.
  • Figure 3A illustrates the thirty-two by thirty-two grid of black or white pixels that make up the limited visual world of the prototype program.
  • the thirty-two by thirty-two grid of black or white pixels presents ample room to create a variety of line drawings and closely related versions of those line drawings that should be identified as belonging to the same group. Examples of some of the line drawings created in the thirty-two by thirty-two black and white pixel grid are presented in Figure 8.
  • the hierarchical arrangement of cortical processing units in the prototype program analyzes the entire thirty-two by thirty-two grid of pixels first during a learning phase and then later during a pattern recognition phase.
  • the lowest level of cortical processing units each process an input sequence of patterns from a four by four patch of pixels.
  • Figure 3B illustrates the thirty-two by thirty-two grid of pixels divided into a set of four by four patches of pixels. Each individual four by four patch of pixels is analyzed by its own first level cortical processing unit. There are sixty-four individual four by four pixel patches illustrated in Figure 3B such that the first cortical processing unit layer of the hierarchy will have sixty-four cortical processing units.
  • Patch 310 illustrates an example four by four patch of pixels taken from the full thirty-two by thirty-two grid of pixels. Each patch of pixels can be addressed using the row and column numbers illustrated on grid of Figure 3B.
  • a full address should also include an identifier of the layer of the cortical processing unit hierarchy layer, hi this example, the lowest hierarchical layer will be deemed layer "X".
  • the address for the example patch could be given as Xl 7 wherein the first character specifies the cortical unit layer, the second character specifies the row, and the third character specifies the column.
  • each cortical processing unit was assigned to examine the sequence of patterns created by four adjacent cortical processing units.
  • Figure 3C illustrates sixteen groups of four by four pixel patches on the grid that are examined by the sixteen cortical processing units of the second cortical layer.
  • the sixteen patches of Figure 3C (and their respective cortical processing units) can be addressed using a row and column index.
  • the corner patch 320 can be addressed as Y03 wherein Y is the layer, 0 is the row, and 3 is the column.
  • FIG. 4 illustrates one possible example of the hierarchy of the cortical processing units for the thirty-two by thirty-two grid of pixels illustrated in Figures 3A to 3C.
  • a third layer has four cortical processing units wherein each processing unit handles data from four cortical processing units in the second layer.
  • a fourth cortical processing unit layer consists of two cortical processing units wherein each considers the outputs from two cortical processing units in the third layer.
  • fifth and top cortical processing unit layer processes the output from the two cortical processing units of the fourth layer.
  • cortical processing units in the successively higher layers of the hierarchy in Figure 4 process information from more than one cortical processing unit in the layer below.
  • the receptive field of successively higher cortical processing units grows larger as one ascends the hierarchy of cortical processing units. This is designed to be similar to the visual system of the human neocortex that operates in a similar manner.
  • Figure 4 presents only one arbitrary selected hierarchical arrangement for the pattern recognition problem. Many other hierarchical arrangements could be created.
  • the processing units in the lowest layer of the hierarchy of Figure 4 each consider an input consisting of sixteen pixels in a four by four pixel patch. But many other divisions of sensory input could be used such as having the cortical units on the lowest layer only consider input from four pixels in two by two pixel patches.
  • the same principle applies to all of the different cortical processing unit layers in the hierarchy.
  • the cortical processing units on the third layer 430 could consider the output from five or six cortical processing units in the second layer.
  • the cortical processing units on the third layer 430 may each only consider the output from four processing units in the lower second layer.
  • FIG. 5A illustrates the cortical hierarchy architecture used to construct the prototype program.
  • the lowest cortical processing unit layer, the X layer 510 has sixty-four cortical processing units for processing the sensory input from the thirty-two by thirty-two pixel grid in individual four by four pixel patches as set forth with Figure 3B.
  • the second cortical processing unit layer, the Y layer 520 has sixteen cortical processing units wherein each cortical processing unit processes the output from four cortical processing units in the X layer 510 as set forth with Figure 3C.
  • a single cortical processing unit comprised the top layer, named the Z layer 530.
  • the single processing unit on the Z layer 530 processes the output from all sixteen cortical processing units on the Y layer 520.
  • Figure 6 illustrates an alternate view of the cortical processing unit hierarchy of Figure 5A.
  • each of the cortical processing units is labeled with an address using the addressing system set forth with reference to Figures 3B and 3C.
  • Figure 8 illustrates an example set of ninety training shapes that were used to train the prototype program.
  • the line drawing shapes of Figure 8 were presented to the prototype program in the form of animated movies that had the line drawing training shape moving about in the thirty-two by thirty-two pixel world. This training allows the prototype program to learn about various geometric shapes in Figure 8 that make up the world of the prototype program and the movements of those geometric shapes in its limited visual world.
  • the cortical processing units of the prototype program each memorize the most frequently occurring sequences of patterns that are encountered. For example, the cortical processing units in the lowest X layer 510 of the prototype program memorizes the most frequently occurring sequences of patterns that are encountered on their local four by four pixel patch. In one embodiment of the prototype program, the cortical processing units in the lowest X layer 510 memorize sequences of four consecutive patterns. Similar pattern sequence learning also occurs at the Y layer and the Z layer. This unsupervised learning allows the prototype program to learn the various primitive structures that make up its world.
  • the unsupervised training was initially performed with animated movies of the specific objects that were to later be recognized.
  • the unsupervised training may have used the letter line drawings in Figure 8.
  • another unsupervised training session used a set of shapes that did not include any of the shapes that the system would later attempt to recognize.
  • the low-level primitive structures in the letters include structures such as vertical lines, horizontal lines, corners, boxes, and parallel lines.
  • Figure 9 illustrates eight different possible sequences of four consecutive four by four pixel patterns that could be seen by the cortical processing units of the lowest layer of the cortical processing unit hierarchy (X layer 510).
  • the four horizontal sequences of patterns Sj, S 2 , S 3 , and S 4 represent four pattern sequences commonly encountered during the animated movies of the unsupervised training.
  • Si and S 2 depict a vertical line moving right and a horizontal line moving upward, respectively.
  • S 3 and S 4 depict the motion of a lower left corner and an upper-right corner, respectively.
  • the four vertical sequences of patterns S 5 , S 6 , S 7 , and S 8 depict very random pattern sequences that never (or almost never) occur.
  • frequently occurring (repeated) sequences of patterns such as those of Si, S 2 , S 3 , and S 4 and other frequently occurring sequences were memorized by the lowest X layer 510 cortical processing units as sequences that contain invariant structure.
  • only the thirteen most frequently occurring sequences of patterns were stored in the memory of cortical processing units in the X layer 510.
  • Each of the memorized frequently occurring sequence of patterns is given a unique identifier.
  • each of the different memorized sequences were very different such that the X layer 510 cortical processing units could immediately recognize which memorized sequence was being experienced. (The sequences had no or almost no shared patterns.)
  • each cortical processing unit in the Y layer 520 receives information from four cortical processing units in the lower X layer 510.
  • the four sequence identifiers from the four X layer units are combined to form a spatial pattern received by the cortical processing unit in the Y layer 520.
  • the sequence identifiers from the four X layer processing units are simply concatenated together to form a unique pattern that is presented to the cortical processing unit in the Y layer 520. Note that the Y layer cortical processing units understand higher level concepts since Y layer processing units combine together sequence information from four individual X layer processing units.
  • one embodiment memorized thirteen different commonly encountered sequences in the X layer 510.
  • the cortical processing units in the Y layer 520 may perform the same learning operation as the cortical processing units in the X layer 510 as set forth above.
  • each cortical processing unit in the Y layer 520 identifies and then memorizes commonly experienced sequences of patterns on its input stream.
  • the Y layer processing units would later attempt to recognize those memorized sequences in their input streams.
  • the Y layer processing units could then report recognized sequences of patterns to the next higher cortical processing unit layer, the Z layer 530.
  • the Z layer 530 receives sequence identifiers from all sixteen different processing units in the Y layer 520.
  • the sixteen sequence identifiers from the Y layer 520 may be combined to form a spatial pattern received by the Z layer 530.
  • the prototype program also collects a set of statistics about the patterns and pattern sequences that are detected during the unsupervised learning phase.
  • the collected statistics may include how many times a particular pattern or sequence is encountered and how many times a particular input value is encountered.
  • one very important statistic to keep track of is the specific context when a particular pattern or sequence encountered.
  • the particular higher-layer pattern may be the context for the lower layer patterns and sequences. It is during this training phase that the system constructs a Bayes network such that Bayesian inference using Bayesian belief propagation may be used to perform pattern recognition. Details on Bayesian inference will be set forth in greater detail in a later section.
  • the prototype program After the unsupervised training, the prototype program has collected sequences of patterns and developed higher level object concepts from the lower layer sequences. Since the prototype program already understands object concepts at the top level, the training from layer Y to layer Z can proceed in a supervised manner. The supervised training links specific object names to object concepts that have been learned at the high Z layer. In this manner, the prototype program may later infer the specific object names when giving novel input data.
  • the supervised training may be performed by presenting an object name to the top Z layer 530 while presenting an animated movie of the particular object shape.
  • the prototype program thus links the high level concept at the Z layer (the current sequence with the object name currently being presented.
  • the memorized sequences recognized at the Z layer are associated with the name "cat”.
  • the memorized sequences recognized at the Z layer are associated with the name "dog”. And so on with all the different geometric shapes during training.
  • the prototype program also collects a set of statistics about the detected patterns and pattern sequences during the supervised learning phase. These statistics will be used to build probability tables in order to perform Bayesian inference as will be described in a later section.
  • the learning pattern recognition prototype program uses Bayesian inference techniques in order to generate very accurate object name using pattern recognition.
  • the objective is to have recognized sequence information at the top Z layer 530 associated with the correct object name.
  • recognized Z layer sequences and the object name associated with the recognized Z layer sequence
  • Bayes' rule allows for a system to determine the probability of a particular result R outcome based upon available evidence E. To compute this probability, one needs to have observed a fair amount of information about the different possible results R and the different possible evidence E.
  • the fundamental rule of Bayesian learning is that P(R
  • E) P(E
  • the prototype program To perform the desired pattern recognition with the prototype program, the prototype program must attempt to determine the output sequence Z at the top of the hierarchy (and thus the associated object name) when given particular input information at the bottom (X layer 510) of the hierarchy. Since this is a bottom to top determination, it is technically referred to as an inference.
  • the input information is supplied in the form of a static thirty-two by thirty-two pixel drawing that is provided into the lowest layer, the X layer 510.
  • the lowest X layer 510 processes the thirty-two by thirty-two pixel drawing to determine the mostly likely sequence at the X layer. If given time, the X layer could also receive and consider real pattern sequence information to generate more accurate results.
  • a Y sequence value can be selected using a probability of the Y sequence value given specific X sequence inputs [P(Y
  • the Z sequence value can be selected using a probability of the Z value given selected Y sequence inputs [P(Z
  • These probabilities may be computed using Bayes' rule. In order to perform this, all of the individual Y layer processing units can observe the probability of a particular Y sequence value given a specific Z sequence value, P(Y
  • Figure 5B illustrates examples of the contextual feedback that may be presented from higher cortical layers to lower cortical layers in order to allow the needed probability tables to be created.
  • the current Z sequence context information is provided for each cortical processing unit in the Y layer. This is illustrated in Figure 5B as the current Z sequence value being fed back to Y layer processing units.
  • This Z sequence contextual feedback allows each cortical processing unit in the Y layer to create a probability table that specifies the probabilities of all the different Y sequences when given a specific Z sequence, P(Y
  • the sixteen probability tables in the sixteen processing units in the Y layer may appear as follows:
  • the current Y sequence context information is provided from its associated Y layer processing unit. This is illustrated in Figure 5B as the current Y sequence value from Y layer processing units YOO being fed back to two associated lower X layer processing units. (Note that only two feed back paths are illustrated from a Y layer processing unit to two X layer processing units.
  • the prototype program is first trained during a training phase using shapes such as the shapes illustrated in Figure 8.
  • shapes such as the shapes illustrated in Figure 8.
  • each of the cortical processing units will extract structure observed in its world in the form of commonly repeated sequences of patterns.
  • Such training may take place unsupervised. Furthermore, such training doesn't even require the specific objects that will later need to be recognized.
  • the training may be occur with any shapes that have the same types of primitive structures that exist in the objects to be recognized.
  • each of the cortical processing units in the prototype program creates memories of the invariant structures that it detects during its exploration of its world.
  • supervised training is performed.
  • an object name associated with a particular object is presented to the prototype program in addition to an animated movie of the object that is being using for training.
  • the prototype program learns to associate the object name with the sequences on the Z layer.
  • novel data with an object image is presented to the prototype program.
  • the prototype program extracts the structure from the object image in the form of stored structure recognized in the X, Y, and Z layers.
  • the recognized structure is used to identify associated sets of probability tables.
  • the identified probability tables are then used to calculate the probabilities of each object name being the correct name of the viewed object.
  • the object name with the highest probability is inferred to be the name of the object in the novel image.
  • Figure 10 illustrates some result data from three different line drawing tests.
  • Column A presents the original training line drawing.
  • Columns B and C illustrate human made line drawings that were successfully identified as belonging to the same class as the object in the Column A.
  • Columns B and C illustrate human made line drawings that were successfully identified as belonging to the same class as the object in the Column A.
  • Column D contains human made line drawings that were not successfully identified as belonging to the same class as the object in the Column A.
  • Column E contains the line drawing class that program believed the line drawing of column D belonged within.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Algebra (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

La présente invention concerne des systèmes à mémoire sophistiqués et des machines intelligentes pouvant être réalisés par création d'un système à mémoire active avec une architecture hiérarchique. Plus spécifiquement, un système peut comprendre plusieurs unités de traitement corticales individuelles disposées de manière à former une structure hiérarchique. Chaque unité de traitement corticale individuelle reçoit une séquence de motifs en tant qu'entrées. Chaque unité de traitement corticale traite la séquence d'entrées reçue qui contient les motifs au moyen d'une mémoire contenant des séquences préalablement rencontrées présentant une structure et elle produit un autre motif. Etant donné que plusieurs séquences d'entrées sont traitées par une unité de traitement corticale, cela permet de produire une séquence de motifs à la sortie. La séquence de motifs à la sortie peut être transmise en tant qu'entrée à une ou plusieurs unités de traitement corticales dans une couche supérieure précédente de la hiérarchie. Une couche inférieure d'unités de traitement corticales peut recevoir une entrée sensorielle depuis le monde extérieur. L'entrée sensorielle comprend également une séquence de motifs.
EP05853611A 2004-12-10 2005-12-09 Procedes, architecture et dispositif permettant de mettre en oeuvre des systemes a memoire hierarchique et a intelligence machine Withdrawn EP1836657A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US11/010,243 US20060184462A1 (en) 2004-12-10 2004-12-10 Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
PCT/US2005/044729 WO2006063291A2 (fr) 2004-12-10 2005-12-09 Procedes, architecture et dispositif permettant de mettre en oeuvre des systemes a memoire hierarchique et a intelligence machine

Publications (1)

Publication Number Publication Date
EP1836657A2 true EP1836657A2 (fr) 2007-09-26

Family

ID=36406007

Family Applications (1)

Application Number Title Priority Date Filing Date
EP05853611A Withdrawn EP1836657A2 (fr) 2004-12-10 2005-12-09 Procedes, architecture et dispositif permettant de mettre en oeuvre des systemes a memoire hierarchique et a intelligence machine

Country Status (6)

Country Link
US (3) US20060184462A1 (fr)
EP (1) EP1836657A2 (fr)
CN (1) CN101107618A (fr)
CA (1) CA2589491A1 (fr)
IL (1) IL183625A0 (fr)
WO (1) WO2006063291A2 (fr)

Families Citing this family (67)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
MY138544A (en) * 2003-06-26 2009-06-30 Neuramatix Sdn Bhd Neural networks with learning and expression capability
US20060184462A1 (en) * 2004-12-10 2006-08-17 Hawkins Jeffrey C Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
US7739208B2 (en) * 2005-06-06 2010-06-15 Numenta, Inc. Trainable hierarchical memory system and method
US20070192267A1 (en) * 2006-02-10 2007-08-16 Numenta, Inc. Architecture of a hierarchical temporal memory based system
US20080208966A1 (en) * 2007-02-28 2008-08-28 Numenta, Inc. Hierarchical Temporal Memory (HTM) System Deployed as Web Service
US8732098B2 (en) 2006-02-10 2014-05-20 Numenta, Inc. Hierarchical temporal memory (HTM) system deployed as web service
US7941389B2 (en) * 2006-02-10 2011-05-10 Numenta, Inc. Hierarchical temporal memory based system including nodes with input or output variables of disparate properties
US9092593B2 (en) 2007-09-25 2015-07-28 Power Analytics Corporation Systems and methods for intuitive modeling of complex networks in a digital environment
US20170046458A1 (en) 2006-02-14 2017-02-16 Power Analytics Corporation Systems and methods for real-time dc microgrid power analytics for mission-critical power systems
US9557723B2 (en) * 2006-07-19 2017-01-31 Power Analytics Corporation Real-time predictive systems for intelligent energy monitoring and management of electrical power networks
US20160246905A1 (en) 2006-02-14 2016-08-25 Power Analytics Corporation Method For Predicting Arc Flash Energy And PPE Category Within A Real-Time Monitoring System
US8131401B2 (en) * 2006-07-19 2012-03-06 Power Analytics Corporation Real-time stability indexing for intelligent energy monitoring and management of electrical power network system
US7937342B2 (en) * 2006-11-28 2011-05-03 Numenta, Inc. Method and apparatus for detecting spatial patterns
US7941392B2 (en) * 2007-02-28 2011-05-10 Numenta, Inc. Scheduling system and method in a hierarchical temporal memory based system
US8112367B2 (en) * 2007-02-28 2012-02-07 Numenta, Inc. Episodic memory with a hierarchical temporal memory based system
WO2008106615A1 (fr) 2007-02-28 2008-09-04 Numenta, Inc. Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques
US20080222064A1 (en) * 2007-03-08 2008-09-11 Larimer Daniel J Processes and Systems for Automated Collective Intelligence
WO2009020684A1 (fr) * 2007-05-16 2009-02-12 Edsa Micro Corporation Systèmes prédictifs en temps réel pour la surveillance d'énergie et la gestion intelligentes de réseaux d'alimentation électrique
US8229722B2 (en) * 2007-05-16 2012-07-24 Power Analytics Corporation Electrical power system modeling, design, analysis, and reporting via a client-server application framework
EP2162853A1 (fr) * 2007-06-29 2010-03-17 Numenta, Inc. Système de mémoire temporelle hiérarchique avec capacité d'inférence améliorée
US8127075B2 (en) * 2007-07-20 2012-02-28 Seagate Technology Llc Non-linear stochastic processing storage device
US8155943B2 (en) * 2007-10-12 2012-04-10 Power Analytics Corporation Systems and methods for automatically converting CAD drawing files into intelligent objects with database connectivity for the design, analysis, and simulation of electrical power systems
WO2009052407A1 (fr) * 2007-10-18 2009-04-23 Numenta, Inc. Système et procédé pour une détermination de topologie automatique dans un réseau temporel hiérarchique
US8175984B2 (en) * 2007-12-05 2012-05-08 Numenta, Inc. Action based learning
US20090216968A1 (en) * 2008-02-27 2009-08-27 Martin Gregory D Method and apparatus for storing sequential sample data as memories for the purpose of rapid memory recognition using mathematic invariants
US8175985B2 (en) 2008-03-19 2012-05-08 Numenta, Inc. Plugin infrastructure for hierarchical temporal memory (HTM) system
US7983998B2 (en) 2008-03-21 2011-07-19 Numenta, Inc. Feedback in group based hierarchical temporal memory system
US8407166B2 (en) * 2008-06-12 2013-03-26 Numenta, Inc. Hierarchical temporal memory system with higher-order temporal pooling capability
US8195582B2 (en) * 2009-01-16 2012-06-05 Numenta, Inc. Supervision based grouping of patterns in hierarchical temporal memory (HTM)
US20110082597A1 (en) 2009-10-01 2011-04-07 Edsa Micro Corporation Microgrid model based automated real time simulation for market based electric power system optimization
EP2548096B1 (fr) * 2010-03-15 2017-10-25 Numenta, Inc. Mémoire temporelle utilisant une représentation distribuée sporadique
US11651277B2 (en) 2010-03-15 2023-05-16 Numenta, Inc. Sparse distributed representation for networked processing in predictive system
US8909950B1 (en) 2010-04-18 2014-12-09 Aptima, Inc. Systems and methods of power management
US8694283B2 (en) * 2010-09-17 2014-04-08 Siemens Aktiengesellschaft System and method for modeling conditional dependence for anomaly detection in machine condition monitoring
US9177259B1 (en) * 2010-11-29 2015-11-03 Aptima Inc. Systems and methods for recognizing and reacting to spatiotemporal patterns
US8645291B2 (en) 2011-08-25 2014-02-04 Numenta, Inc. Encoding of data for processing in a spatial and temporal memory system
US8825565B2 (en) 2011-08-25 2014-09-02 Numenta, Inc. Assessing performance in a spatial and temporal memory system
US8504570B2 (en) 2011-08-25 2013-08-06 Numenta, Inc. Automated search for detecting patterns and sequences in data using a spatial and temporal memory system
US9336774B1 (en) 2012-04-20 2016-05-10 Google Inc. Pattern recognizing engine
US9159021B2 (en) 2012-10-23 2015-10-13 Numenta, Inc. Performing multistep prediction using spatial and temporal memory system
US9904889B2 (en) 2012-12-05 2018-02-27 Applied Brain Research Inc. Methods and systems for artificial cognition
US9749253B2 (en) * 2013-03-14 2017-08-29 Silver Spring Networks, Inc. Technique for implementing a latency sensitive communication protocol in a wireless mesh network
US10534995B2 (en) 2013-03-15 2020-01-14 Qylur Intelligent Systems, Inc. Network of intelligent machines
WO2014149070A1 (fr) * 2013-03-15 2014-09-25 Hrl Laboratories, Llc Réseau de neurones et procédé de programmation
US20160154631A1 (en) * 2013-07-12 2016-06-02 Bryant G. CRUSE Method and system for machine comprehension
EP3028201A4 (fr) * 2013-08-02 2017-03-22 Auckland Uniservices Limited Système d'animation neuro-comportemental
US10318878B2 (en) 2014-03-19 2019-06-11 Numenta, Inc. Temporal processing scheme and sensorimotor information processing
US10817785B2 (en) * 2014-08-09 2020-10-27 Fred Narcross Brain emulator support system
US9886193B2 (en) * 2015-05-15 2018-02-06 International Business Machines Corporation Architecture and implementation of cortical system, and fabricating an architecture using 3D wafer scale integration
US11100414B2 (en) 2016-05-13 2021-08-24 Numenta, Inc. Inferencing and learning based on sensorimotor input data
WO2017197375A1 (fr) * 2016-05-13 2017-11-16 Princeton University Système et procédés pour faciliter la reconnaissance de formes
WO2017197298A1 (fr) * 2016-05-13 2017-11-16 Numenta, Inc. Déduction et apprentissage basés sur des données d'entrée sensorimotrices
US9792889B1 (en) * 2016-11-03 2017-10-17 International Business Machines Corporation Music modeling
US10769447B1 (en) * 2016-12-01 2020-09-08 Marc Joseph Kirch Synchronous convergent pulses reduce dimensions of serial surfaces
US20190065935A1 (en) * 2017-08-30 2019-02-28 International Business Machines Corporation Computational method for feedback in a hierarchical neural network
CN108107729B (zh) * 2017-12-15 2019-11-08 东华大学 基于类脑分层记忆机制的实时数据流智能化处理平台
US10970080B2 (en) 2018-02-08 2021-04-06 Marvell Asia Pte, Ltd. Systems and methods for programmable hardware architecture for machine learning
US11016801B1 (en) 2018-05-22 2021-05-25 Marvell Asia Pte, Ltd. Architecture to support color scheme-based synchronization for machine learning
US10929760B1 (en) 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Architecture for table-based mathematical operations for inference acceleration in machine learning
US10929779B1 (en) * 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Architecture to support synchronization between core and inference engine for machine learning
US10929778B1 (en) * 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Address interleaving for machine learning
US10997510B1 (en) 2018-05-22 2021-05-04 Marvell Asia Pte, Ltd. Architecture to support tanh and sigmoid operations for inference acceleration in machine learning
US10891136B1 (en) 2018-05-22 2021-01-12 Marvell Asia Pte, Ltd. Data transmission between memory and on chip memory of inference engine for machine learning via a single data gathering instruction
US11205137B2 (en) 2018-06-15 2021-12-21 International Business Machines Corporation Distributed training in a parameter dataset
US10878321B2 (en) * 2019-02-28 2020-12-29 DeepCube LTD. Partial activation of multiple pathways in neural networks
US11681922B2 (en) 2019-11-26 2023-06-20 Numenta, Inc. Performing inference and training using sparse neural network
US11270127B1 (en) 2021-05-05 2022-03-08 Marc Joseph Kirch Synchronized pulses identify and locate targets rapidly

Family Cites Families (81)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4766534A (en) 1986-10-16 1988-08-23 American Telephone And Telegraph Company, At&T Bell Laboratories Parallel processing network and method
US4845744A (en) 1986-10-16 1989-07-04 American Telephone And Telegraph Company, At&T Bell Laboratories Method of overlaying virtual tree networks onto a message passing parallel processing network
US5255348A (en) * 1991-06-14 1993-10-19 Nenov Valeriy I Neural network for learning, recognition and recall of pattern sequences
US7481453B2 (en) 1991-07-09 2009-01-27 Automotive Technologies International, Inc. Inflator system
US6400996B1 (en) 1999-02-01 2002-06-04 Steven M. Hoffberg Adaptive pattern recognition based control system and method
US5729661A (en) 1992-11-24 1998-03-17 Pavilion Technologies, Inc. Method and apparatus for preprocessing input data to a neural network
US7251637B1 (en) 1993-09-20 2007-07-31 Fair Isaac Corporation Context vector generation and retrieval
JPH0877010A (ja) 1994-09-07 1996-03-22 Hitachi Ltd データ分析方法および装置
US7786864B1 (en) 2000-09-08 2010-08-31 Automotive Technologies International, Inc. Vehicular RFID and sensor assemblies
US5712953A (en) * 1995-06-28 1998-01-27 Electronic Data Systems Corporation System and method for classification of audio or audio/video signals based on musical content
AU4238697A (en) 1996-08-29 1998-03-19 Cisco Technology, Inc. Spatio-temporal processing for communication
US6366896B1 (en) 1997-03-18 2002-04-02 William R. Hutchison Adaptive agent using neural network
US6028608A (en) * 1997-05-09 2000-02-22 Jenkins; Barry System and method of perception-based image generation and encoding
US6516416B2 (en) 1997-06-11 2003-02-04 Prism Resources Subscription access system for use with an untrusted network
US6353814B1 (en) 1997-10-08 2002-03-05 Michigan State University Developmental learning machine and method
US6195622B1 (en) * 1998-01-15 2001-02-27 Microsoft Corporation Methods and apparatus for building attribute transition probability models for use in pre-fetching resources
US6636631B2 (en) 1998-06-04 2003-10-21 Matsushita Electric Industrial Co., Ltd. Optical character reading method and system for a document with ruled lines and its application
US6567814B1 (en) 1998-08-26 2003-05-20 Thinkanalytics Ltd Method and apparatus for knowledge discovery in databases
US6122014A (en) 1998-09-17 2000-09-19 Motorola, Inc. Modified chroma keyed technique for simple shape coding for digital video
US6882992B1 (en) 1999-09-02 2005-04-19 Paul J. Werbos Neural networks for intelligent control
US6751343B1 (en) * 1999-09-20 2004-06-15 Ut-Battelle, Llc Method for indexing and retrieving manufacturing-specific digital imagery based on image content
US6468069B2 (en) * 1999-10-25 2002-10-22 Jerome H. Lemelson Automatically optimized combustion control
US6625585B1 (en) 2000-02-18 2003-09-23 Bioreason, Inc. Method and system for artificial intelligence directed lead discovery though multi-domain agglomerative clustering
US6917845B2 (en) 2000-03-10 2005-07-12 Smiths Detection-Pasadena, Inc. Method for monitoring environmental condition using a mathematical model
NZ503882A (en) 2000-04-10 2002-11-26 Univ Otago Artificial intelligence system comprising a neural network with an adaptive component arranged to aggregate rule nodes
US20020006121A1 (en) 2000-04-27 2002-01-17 Dileep George Adaptive diversity combining for wide band code division multiple access (W-CDMA) based on iterative channel estimation
US7054850B2 (en) 2000-06-16 2006-05-30 Canon Kabushiki Kaisha Apparatus and method for detecting or recognizing pattern by employing a plurality of feature detecting elements
EP1295252B1 (fr) 2000-06-30 2006-12-27 BRITISH TELECOMMUNICATIONS public limited company Appareil pour generer des sequences d'elements
US6714941B1 (en) 2000-07-19 2004-03-30 University Of Southern California Learning data prototypes for information extraction
DE60142582D1 (de) 2000-10-13 2010-08-26 Fraunhofer Ges Forschung Verfahren zum überwachten trainieren eines iterativen künstlichen neuronalen netzwerks
WO2002063402A1 (fr) 2001-02-05 2002-08-15 Honda Giken Kogyo Kabushiki Kaisha Appareil, procede et programme d'apprentissage pour agent
US6574537B2 (en) 2001-02-05 2003-06-03 The Boeing Company Diagnostic system and method
US6947378B2 (en) 2001-02-28 2005-09-20 Mitsubishi Electric Research Labs, Inc. Dynamic network resource allocation using multimedia content features and traffic features
US6615211B2 (en) 2001-03-19 2003-09-02 International Business Machines Corporation System and methods for using continuous optimization for ordering categorical data sets
US7962482B2 (en) 2001-05-16 2011-06-14 Pandora Media, Inc. Methods and systems for utilizing contextual feedback to generate and modify playlists
US7797062B2 (en) * 2001-08-10 2010-09-14 Rockwell Automation Technologies, Inc. System and method for dynamic multi-objective optimization of machine selection, integration and utilization
US20040142325A1 (en) * 2001-09-14 2004-07-22 Liat Mintz Methods and systems for annotating biomolecular sequences
US7233781B2 (en) 2001-10-10 2007-06-19 Ochoa Optics Llc System and method for emergency notification content delivery
US6957241B2 (en) 2002-02-14 2005-10-18 Gallitzin Allegheny Llc FFT and FHT engine
US7035431B2 (en) 2002-02-22 2006-04-25 Microsoft Corporation System and method for probabilistic exemplar-based pattern tracking
US7203635B2 (en) 2002-06-27 2007-04-10 Microsoft Corporation Layered models for context awareness
US20060098886A1 (en) 2003-01-10 2006-05-11 Koninklijke Philips Electronics N.V. Efficient predictive image parameter estimation
US20040148520A1 (en) 2003-01-29 2004-07-29 Rajesh Talpade Mitigating denial of service attacks
US7627171B2 (en) 2003-07-03 2009-12-01 Videoiq, Inc. Methods and systems for detecting objects of interest in spatio-temporal signals
US8510083B2 (en) 2003-07-31 2013-08-13 The Boeing Company Method, apparatus and computer program product for constructing a diagnostic network model
JP3987013B2 (ja) * 2003-09-01 2007-10-03 本田技研工業株式会社 車両周辺監視装置
US7565451B2 (en) 2004-01-23 2009-07-21 Microsoft Corporation Adaptive dispatch of received messages to code using inter-positioned message modification
WO2005072431A2 (fr) 2004-01-27 2005-08-11 Sarnoff Corporation Procede et appareil de combinaison de plusieurs images
US7853456B2 (en) 2004-03-05 2010-12-14 Health Outcomes Sciences, Llc Systems and methods for risk stratification of patient populations
US7788109B2 (en) 2004-04-03 2010-08-31 Altusys Corp. Method and apparatus for context-sensitive event correlation with external control in situation-based management
US7565194B2 (en) 2004-05-12 2009-07-21 Zoll Medical Corporation ECG rhythm advisory method
WO2006137829A2 (fr) 2004-08-10 2006-12-28 Sarnoff Corporation Procede et systeme d'execution d'acquisition d'image adaptative
US20060184462A1 (en) 2004-12-10 2006-08-17 Hawkins Jeffrey C Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
TWI273409B (en) * 2005-01-18 2007-02-11 Sunplus Technology Co Ltd Pre-fetch technology by read merge mechanism
JP2006285899A (ja) 2005-04-05 2006-10-19 Sony Corp 学習装置および学習方法、生成装置および生成方法、並びにプログラム
US7577651B2 (en) 2005-04-28 2009-08-18 Yahoo! Inc. System and method for providing temporal search results in response to a search query
US7809722B2 (en) 2005-05-09 2010-10-05 Like.Com System and method for enabling search and retrieval from image files based on recognized information
US7739208B2 (en) 2005-06-06 2010-06-15 Numenta, Inc. Trainable hierarchical memory system and method
CA2573855A1 (fr) * 2006-01-25 2007-07-25 Espeed, Inc. Systemes et methode de facilitation de completion d'ententes de rachat
US20070192267A1 (en) 2006-02-10 2007-08-16 Numenta, Inc. Architecture of a hierarchical temporal memory based system
US7941389B2 (en) 2006-02-10 2011-05-10 Numenta, Inc. Hierarchical temporal memory based system including nodes with input or output variables of disparate properties
US20080208966A1 (en) * 2007-02-28 2008-08-28 Numenta, Inc. Hierarchical Temporal Memory (HTM) System Deployed as Web Service
US8732098B2 (en) 2006-02-10 2014-05-20 Numenta, Inc. Hierarchical temporal memory (HTM) system deployed as web service
US7840395B2 (en) 2006-03-10 2010-11-23 Edsa Micro Corporation Systems and methods for predictive monitoring including real-time strength and security analysis in an electrical power distribution system
US7826990B2 (en) 2006-02-14 2010-11-02 Edsa Micro Corporation Systems and methods for real-time system monitoring and predictive analysis
CA2646423A1 (fr) 2006-03-10 2007-09-20 Edsa Micro Corporation Systemes et procedes d'evaluation de dispositifs de protection en temps reel dans un systeme de distribution d'energie electrique
US7840396B2 (en) 2006-03-10 2010-11-23 Edsa Micro Corporation Systems and methods for determining protective device clearing times used for providing real-time predictions about arc flash events
US7844440B2 (en) 2006-07-07 2010-11-30 Edsa Micro Corporation Systems and methods for real-time dynamic simulation of uninterruptible power supply solutions and their control logic systems
US7937342B2 (en) 2006-11-28 2011-05-03 Numenta, Inc. Method and apparatus for detecting spatial patterns
WO2008106615A1 (fr) 2007-02-28 2008-09-04 Numenta, Inc. Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques
US7941392B2 (en) * 2007-02-28 2011-05-10 Numenta, Inc. Scheduling system and method in a hierarchical temporal memory based system
US8112367B2 (en) * 2007-02-28 2012-02-07 Numenta, Inc. Episodic memory with a hierarchical temporal memory based system
EP2162853A1 (fr) 2007-06-29 2010-03-17 Numenta, Inc. Système de mémoire temporelle hiérarchique avec capacité d'inférence améliorée
US7676458B2 (en) 2007-08-28 2010-03-09 International Business Machines Corporation System and method for historical diagnosis of sensor networks
WO2009052407A1 (fr) 2007-10-18 2009-04-23 Numenta, Inc. Système et procédé pour une détermination de topologie automatique dans un réseau temporel hiérarchique
US8175984B2 (en) 2007-12-05 2012-05-08 Numenta, Inc. Action based learning
US8175985B2 (en) 2008-03-19 2012-05-08 Numenta, Inc. Plugin infrastructure for hierarchical temporal memory (HTM) system
US7983998B2 (en) 2008-03-21 2011-07-19 Numenta, Inc. Feedback in group based hierarchical temporal memory system
US8407166B2 (en) 2008-06-12 2013-03-26 Numenta, Inc. Hierarchical temporal memory system with higher-order temporal pooling capability
US8195582B2 (en) 2009-01-16 2012-06-05 Numenta, Inc. Supervision based grouping of patterns in hierarchical temporal memory (HTM)
EP2548096B1 (fr) 2010-03-15 2017-10-25 Numenta, Inc. Mémoire temporelle utilisant une représentation distribuée sporadique

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See references of WO2006063291A2 *

Also Published As

Publication number Publication date
WO2006063291A2 (fr) 2006-06-15
US9530091B2 (en) 2016-12-27
US20060184462A1 (en) 2006-08-17
US8175981B2 (en) 2012-05-08
CA2589491A1 (fr) 2006-06-15
CN101107618A (zh) 2008-01-16
US20080201286A1 (en) 2008-08-21
WO2006063291A3 (fr) 2007-06-28
IL183625A0 (en) 2008-01-20
US20120197823A1 (en) 2012-08-02

Similar Documents

Publication Publication Date Title
US8175981B2 (en) Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
Gallant Neural network learning and expert systems
Hawkins et al. On intelligence
Marcus Can connectionism save constructivism?
Harnad et al. Categorical perception and the evolution of supervised learning in neural nets
AU2009202547A1 (en) Neural networks with learning and expression capability
Baxter The evolution of learning algorithms for artificial neural networks
US11132598B1 (en) System and method for humanoid robot control and cognitive self-improvement without programming
EP0526504B1 (fr) Dispositif neural artificiel
Weng Conscious intelligence requires developmental autonomous programming for general purposes
CN116805158A (zh) 生物大脑中意识学习的发展网络模型
US20200257503A1 (en) Auto-Programming for General Purposes and Auto-Programming Operating Systems
Calitoiu et al. Desynchronizing a chaotic pattern recognition neural network to model inaccurate perception
Parisi et al. A unified simulation scenario for language development, evolution and historical change
Zins Neuromorphic Computing Applications in Robotics
Weng Conscious intelligence requires lifelong autonomous programming for general purposes
Leow et al. Visual schemas in neural networks for object recognition and scene analysis
Ahmed et al. Modeling of Goal-oriented Human Motion Evolution using Hidden Markov Models.
Aprile et al. Enaction and Enactive Interfaces: A Handbook of Terms
MacLennan Neural networks, learning, and intelligence.
Stewart Learning in Artificial Life: Conditioning, Concept Formation, and Sensorimotor Loops
Richardson From Neurons to Cognition
Jones Neural Networks and the Computational Brain
Kale UNRAVELING THE COMPLEXITY OF THE HUMAN BRAIN: A MULTIDISCIPLINARY APPROACH TO UNDERSTANDING NEURAL MECHANISMS
Uhr Forms structure form at ever “higher” and “lower” levels

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20070710

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU LV MC NL PL PT RO SE SI SK TR

AX Request for extension of the european patent

Extension state: AL BA HR MK YU

DAX Request for extension of the european patent (deleted)
17Q First examination report despatched

Effective date: 20080623

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20080104